A Micro-Benchmark Suite for Evaluating Hadoop MapReduce …...Hadoop MapReduce 5 Performance of...

Dip$ Shankar, Xiaoyi Lu, Md. Wasi-‐ur-‐Rahman, Nusrat Islam, and Dhabaleswar K. (DK) Panda

Network-‐Based Compu2ng Laboratory

Department of Computer Science and Engineering The Ohio State University, Columbus, OH, USA

A Micro-Benchmark Suite for Evaluating Hadoop MapReduce on High-Performance Networks

BPOE-5 2014

Outline

•  Introduc$on & Mo$va$on

•  Design Considera$ons •  Micro-‐benchmark Suite

•  Performance Evalua$on

•  Case Study with RDMA

•  Conclusion & Future work

2

BPOE-5 2014

Big Data Technology - Hadoop

Hadoop Distributed File System (HDFS)

MapReduce (Cluster Resource Management & Data

Processing)

Hadoop Common/Core (RPC, ..)

Hadoop Distributed File System (HDFS)

YARN (Cluster Resource Management & Job

Scheduling)

Hadoop Common/Core (RPC, ..)

MapReduce (Data Processing)

Other Models (Data Processing)

Hadoop 1.x Hadoop 2.x

•  Apache Hadoop is one of the most popular Big Data technology –  Provides frameworks for large-scale, distributed data storage and

processing –  MapReduce, HDFS, YARN, RPC, etc.

3

BPOE-5 2014 4

Big Data Technology - MapReduce

Disk Operations •  Map and Reduce Tasks carry out the total job execution

–  Map tasks read from HDFS, operate on it, and write the intermediate data to local disk –  Reduce tasks get these data by shuffle from TaskTrackers, operate on it and write to

HDFS •  Scalable and communication intensive

–  Data shuffling

Bulk Data Transfer

BPOE-5 2014

Factors Effecting Performance of Hadoop MapReduce

5

Performance of Hadoop MapReduce is influenced by many factors •  Network configuration of cluster •  Multi-core architecture •  Memory system •  Underlying storage system

–  Example: HDFS, Lustre etc.

•  Controllable parameters in software –  Number of Mappers and Reducers, Partitioning scheme used

•  Many others …

BPOE-5 2014

Common Protocols using Open Fabrics

6 WBDB 2013

Common"Protocols"using"Open"Fabrics"

�

Applica-on&

Verbs&Sockets&

ApplicaAon"Interface"

SDP

RDMA&

SDP&

InfiniBand&Adapter&

InfiniBand&Switch&

RDMA&

IB&Verbs&

InfiniBand&Adapter&

InfiniBand&Switch&

User space

RDMA&

RoCE&

RoCE&Adapter&

User space

Ethernet&Switch&

TCP/IP&

Ethernet&Driver&

Kernel Space

Protocol"

InfiniBand&Adapter&

InfiniBand&Switch&

IPoIB&

Ethernet&Adapter&

Ethernet&Switch&

Adapter"

Switch"

1/10/40&

GigE&

iWARP&

Ethernet&Switch&

iWARP&

iWARP&Adapter&

User space IPoIB&

TCP/IP&

Ethernet&Adapter&

Ethernet&Switch&

10/40&GigEL

TOE&

Hardware"Offload"

RSockets&

InfiniBand&Adapter&

InfiniBand&Switch&

User space

RSockets&

BPOE-5 2014

Can High-Performance Networks Benefit Big Data Processing? •  Previous studies: very good performance improvements for Hadoop (HDFS

/MapReduce/RPC), Spark, HBase, Memcached over InfiniBand/RoCE –  Hadoop Acceleration with RDMA

•  N. S. Islam, et.al., SOR-HDFS: A SEDA-based Approach to Maximize Overlapping in RDMA-Enhanced HDFS, HPDC’14

•  N. S. Islam, et.al., High Performance RDMA-Based Design of HDFS over InfiniBand, SC’12 •  M. W. Rahman, et.al. HOMR: A Hybrid Approach to Exploit Maximum Overlapping in

MapReduce over High Performance Interconnects, ICS’14 •  M. W. Rahman, et.al., High-Performance RDMA-based Design of Hadoop MapReduce over

InfiniBand, HPDIC’13 •  X. Lu, et. al., High-Performance Design of Hadoop RPC with RDMA over InfiniBand, ICPP’13

–  Spark Acceleration with RDMA •  X. Lu, et. al., Accelerating Spark with RDMA for Big Data Processing: Early Experiences,

HotI'14

–  HBase Acceleration with RDMA •  J. Huang, et.al., High-Performance Design of HBase with RDMA over InfiniBand, IPDPS’12

–  Memcached Acceleration with RDMA •  J. Jose, et. al., Scalable Memcached design for InfiniBand Clusters using Hybrid Transports,

Cluster’11 •  J. Jose, et.al., Memcached Design on High Performance RDMA Capable Interconnects,

ICPP’11 7

BPOE-5 2014

•  RDMA for Apache Hadoop 2.x (RDMA-Hadoop-2.x)

•  RDMA for Apache Hadoop 1.x (RDMA-Hadoop)

•  RDMA for Memcached (RDMA-Memcached)

•  OSU HiBD-Benchmarks (OHB)

•  http://hibd.cse.ohio-state.edu

•  RDMA for Apache HBase and Spark

The High-Performance Big Data (HiBD) Project

8

BPOE-5 2014

RDMA for Apache Hadoop 1.x/2.x Distributions

http://hibd.cse.ohio-state.edu

•  High-Performance Design of Hadoop over RDMA-enabled Interconnects

–  High performance design with native InfiniBand and RoCE support at the verbs-level for HDFS, MapReduce, and RPC components

–  Easily configurable for native InfiniBand, RoCE and the traditional sockets-based support (Ethernet and InfiniBand with IPoIB)

•  Current release: 0.9.9 (03/31/14) –  Based on Apache Hadoop 1.2.1

–  Compliant with Apache Hadoop 1.2.1 APIs and applications

–  Tested with

•  Mellanox InfiniBand adapters (DDR, QDR and FDR)

•  RoCE support with Mellanox adapters

•  Various multi-core platforms, different file systems with disks and SSDs

•  RDMA for Apache Hadoop 2.x 0.9.1 is released in HiBD!

9

BPOE-5 2014 HPDIC 2013

MapReduce RDMA Design Overview

• Enables high performance RDMA communication, while supporting traditional socket interface

• JNI Layer bridges Java based MapReduce with communication library (UCR) written in native code

10

MapReduce

IB Verbs

RDMA Capable Networks

(IB, 10GE/ iWARP, RoCE ..)

OSU-IB Design

Applications

1/10 GigE Network

Java Socket Interface

Java Native Interface (JNI)

Job Tracker

Task Tracker

Map

Reduce

Design features for RDMA

¾ Prefetch/Caching of MOF

¾ In-Memory Merge

¾ Overlap of Merge &Reduce

Design Overview of MapReduce with RDMA •  Enables high performance RDMA

communication, while supporting traditional socket interface

•  JNI Layer bridges Java based MapReduce with communication library written in native code

•  Design features –  RDMA-based shuffle –  Prefetching and caching map output –  Efficient Shuffle Algorithms –  In-memory merge –  On-demand Shuffle Adjustment –  Advanced overlapping

•  map, shuffle, and merge •  shuffle, merge, and reduce

–  On-demand connection setup –  InfiniBand/RoCE support

M. Wasi-ur-Rahman et al., HOMR: A Hybrid Approach to Exploit Maximum Overlapping in MapReduce over High Performance Interconnects, ICS’14

M. Wasi-ur-Rahman et al., High-Performance RDMA-based Design of Hadoop MapReduce over InfiniBand, HPDIC'13 10

BPOE-5 2014

Existing Benchmarks for Evaluating Hadoop MapReduce

11

Evaluation of performance Hadoop MapReduce job Micro-benchmarks Description

Sort Sort random data, I/O bound

TeraSort Sort in total order, Map Stage is CPU bound, Reduce Stage is I/O bound

Wordcount CPU bound, Heavy use of partitioner

Other Benchmarks Description HiBench Representative and comprehensive suite with both

synthetic micro-benchmarks and real-world workloads

MRBS Five benchmarks covering several application domains for evaluating the dependability of MapReduce systems

MRBench Focuses on processing business oriented queries and concurrent data modifications

PUMA Represents broad range of MapReduce applications with high/low computation and high/low shuffle volumes

SWIM Real life MapReduce workloads from production systems, workload synthesis and replay tools for sampling traces

BPOE-5 2014

Existing Benchmarks for Evaluating Hadoop MapReduce (contd.)

12

•  All benchmarks mentioned require the involvement of HDFS or some distributed file system –  Interferes in the evaluation of the MapReduce’s performance

–  Hard to benchmark and compare different shuffle and sort schemes

–  Benchmarks do not provision us to study different shuffle patterns and impact of network protocols on them

•  Requirement: we need a way to evaluate MapReduce as an independent component ! –  To illustrate performance improvement and potential of new

MapReduce designs

–  To help tune internal parameters specific to MapReduce and obtain optimal performance over high-performance networks

BPOE-5 2014

Can we benefit from a stand-alone MapReduce Micro-benchmark?

13

•  Can we design a simple micro-benchmark suite –  That lets users and developers evaluate Hadoop MapReduce

in a stand-alone manner over different networks or protocols? –  Helps tune and optimize configurable parameters based on

cluster and workload characteristics? –  Helps us evaluate the performance of new or alternate

MapReduce frameworks such as our proposed MapReduce over RDMA design?

–  That provisions studying the impact of different data distribution patterns, data types, etc., on the performance of the MapReduce job?

•  What will be the performance of stand-alone Hadoop MapReduce when evaluated on different high-performance networks?

•  Basic motivation!

BPOE-5 2014

Outline






14

BPOE-5 2014

,QWHUPHGLDWH�'DWD�

'LVWULEXWLRQ 6L]H�DQG�QXPEHU�RI�NH\�YDOXH�SDLUV

1XPEHU�RI�PDS�DQG�

UHGXFH�WDVNV

'DWD�7\SH�RI�NH\�YDOXH�SDLUV

1HWZRUN�&RQILJXUDWLRQ

5HVRXUFH�8WLOL]DWLRQ�

�&38�QHWZRUN�

Design Considerations

15

•  Intermediate data distribution –  Is the shuffle data partitioned

evenly amongst all reducers?

•  Size and number of key/value pairs –  Does the size of intermediate

data matter?

–  Size Vs. Number of key/value pairs

•  Number of map and reduce tasks –  Number of tasks generating

processing the intermediate data

BPOE-5 2014

Design Considerations (contd.)

•  Data type of key/value pairs –  E.g. Bytes Vs. Text

•  Network Configuration –  Help to evaluate performance

of different networks for MapReduce workloads

–  E.g. 1GigE/10GigE/IB

•  Resource Utilization –  Correlation between workload

characteristics and resource utilization in MapReduce

16

,QWHUPHGLDWH�'DWD�

'LVWULEXWLRQ 6L]H�DQG�QXPEHU�RI�NH\�YDOXH�SDLUV

1XPEHU�RI�PDS�DQG�

UHGXFH�WDVNV

'DWD�7\SH�RI�NH\�YDOXH�SDLUV

1HWZRUN�&RQILJXUDWLRQ

5HVRXUFH�8WLOL]DWLRQ�

�&38�QHWZRUN�

BPOE-5 2014

Outline






17

BPOE-5 2014

MapReduce Micro-Benchmark Suite

18

•  Evaluate the performance of stand-alone MapReduce –  Does not require or involve HDFS or any distributed file

system

•  Considers various factors that influence the data shuffling phase –  Underlying network configuration, number of map and reduce tasks,

intermediate shuffle data pattern, shuffle data size etc.

•  Currently supports three different shuffle data distribution patterns –  Average data distribution: intermediate data is evenly distributed

among reduce tasks –  Random data distribution: intermediate data is pseudo-randomly

distributed among reduce tasks –  Skewed data distribution: intermediate data is unevenly

distributed among reduce tasks

BPOE-5 2014 19

Micro-Benchmark Suite Design

&XVWRP�3DUWLWLRQHU

6KXIIOH�DQG�6RUW

5HGXFH

3HUIRUPDQFH�FKDUDFWHULVWLFV

'XPP\�6SOLW

8VHU�VSHFLILHG�SDUDPHWHUV

&XVWRP�3DUWLWLRQHU

6KXIIOH�DQG�6RUW

'XPP\�6SOLW

'XPP\�6SOLW

'XPP\�6SOLW

0DS0DS0DS 0DS

2YHUDOO�3HUIRUPDQFH�5HVXOWV

5HGXFH

3HUIRUPDQFH�FKDUDFWHULVWLFV

'DWD�6KXIIOLQJ3KDVH

*HQHUDWHG�LQWHUPHGLDWH��N�Y��SDLUV

NH\�YDOXH�H[FKDQJH�SDWWHUQ

•  Stand-alone feature –  Simple No-HDFS MapReduce job –  Map Phase

–  Custom Input Format i.e., NullInputFormat

–  key/value pairs generated in memory

–  Reduce Phase uses Hadoop API’s NullOutputFormat

•  Custom Partitioners –  Simulate different intermediate data

distribution scenarios described •  Configurable Parameters

–  Number of maps and reducers –  Intermediate shuffle data size

–  Number of key/value pairs per map –  Size of key/value pairs

–  Data type •  Calculates statistics like job latency,

CPU/Network utilization

BPOE-5 2014

MapReduce Micro-Benchmarks

20

Three MapReduce Micro-benchmarks defined 1) MR-AVG micro-benchmark •  Intermediate data is evenly distributed among reduce tasks •  Custom partitioner distributes same number of intermediate

key/value pairs amongst the reducers in a round-robin fashion

•  Uniform distribution and fair comparison for all runs 2) MR-RAND micro-benchmark •  Intermediate data is pseudo-randomly distributed among

reduce tasks •  Custom partitioner picks a reducer randomly and assigns the

key/value pair to it •  Fair comparison for all runs on homogenous systems

BPOE-5 2014

MapReduce Micro-Benchmarks (contd.)

21

3) MR-SKEW micro-benchmark •  Intermediate data is unevenly distributed among reduce

tasks •  Custom partitioner uses fixed skewed distribution

pattern –  50% to reducer 0 –  25% to reducer 1 –  12.5% to reducer 2 –  …

•  Skewed distribution pattern is fixed for all runs •  Fair comparison for all runs on homogenous systems

BPOE-5 2014

Outline






22

BPOE-5 2014

Experimental Setup •  Hardware

–  Intel Westmere Cluster (A) (Up to 9 nodes) •  Each node has 8 processor cores on 2 Intel Xeon 2.67 GHz

quad-core CPUs, 24 GB main memory •  Mellanox QDR HCAs (32 Gbps) + 10GigE

–  TACC Stampede Cluster (B) (Up to 17 nodes) •  Intel Sandy Bridge (E5-2680) dual octa-core processors,

running at 2.70GHz, 32 GB main memory •  Mellanox FDR HCAs (56 Gbps)

–  Performance comparisons over 1 GigE, 10 GigE, IPoIB (32 Gbps) and, IPoIB (56 Gbps)

•  Software –  JDK 1.7.0 –  Apache Hadoop 1.2.1 –  Apache Hadoop NextGen MapReduce (YARN) 2.4.1

23

BPOE-5 2014

Performance Evaluation with Different Data Distribution Patterns

0

500

1,000

1,500

2,000

2,500

3,000

3,500

4,000

16 32 64 128

Job

Exe

cutio

n T

ime(

seco

nds)

Shuffle Data Size(GB)

1GigE10GigEIPoIB (32Gbps)

0 500

1,000 1,500 2,000 2,500 3,000 3,500 4,000 4,500

16 32 64 128

Job

Exe

cutio

n T

ime(

seco

nds)



24

•  Comparing different shuffle data distribution patterns on Cluster A, key/value pair size of 1 KB on 4 node cluster,16 maps and 8 reduces

•  For MR-AVG, IPoIB (32Gbps) gives a 24% improvement over 1 GigE and up to 17% over 10 GigE

•  For MR-RAND, IPoIB (32Gbps) gives a 22% improvement over 1 GigE and up to 15% over 10 GigE

MR-‐AVG micro-‐benchmark MR-‐RAND micro-‐benchmark

BPOE-5 2014

Performance Evaluation with Different Data Distribution Patterns

•  IPoIB (32Gbps) provides better improvement with increased shuffle data sizes and more skewed workloads

•  Skewed data distribution causes 2x increase in job execution time for a given data size, irrespective of the underlying network interconnect

0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 8,000 9,000

16 32 64 128

Job

Exe

cutio

n T

ime(

seco

nds)



MR-‐SKEW micro-‐benchmark

25

•  For MR-SKEW, 10 GigE gives up to 13% improvement over 1 GigE

•  For MR-SKEW, IPoIB (32Gbps) gives up to 14% improvement over 10 GigE

BPOE-5 2014

Performance Evaluation with Apache Hadoop NextGen MapReduce (YARN)

0 100 200 300 400 500 600 700 800 900

8 16 32 64 128

Job

Exec

utio

n Ti

me(

seco

nds)



0 100 200 300 400 500 600 700 800 900

8 16 32 64 128

Job

Exec

utio

n Ti

me(

seco

nds)



26

•  Running Hadoop 2.4.1 with different data distributions on Cluster A, key/value pair size of 1 KB on 8 node cluster, 32 maps and 16 reducers

•  For MR-AVG, IPoIB (32Gbps) gives a 21% improvement over 1 GigE and up to 12% over 10 GigE

•  For MR-RAND, IPoIB (32Gbps) gives a 19% improvement over 1 GigE and up to 11% over 10 GigE

MR-‐AVG micro-‐benchmark MR-‐RAND micro-‐benchmark

BPOE-5 2014

Performance Evaluation with Apache Hadoop NextGen MapReduce (YARN)

•  Skewed data distribution

–  causes around 3x increase in job execution time for a given data size

–  Increased concurrency does not show significant improvement as Reduce phase still depends on the slowest reduce task

0

1,000

2,000

3,000

4,000

5,000

6,000

8 16 32 64 128

Job

Exe

cutio

n T

ime(

seco

nds)



MR-‐SKEW micro-‐benchmark

27

•  For MR-SKEW, 10 GigE gives up to 12% improvement over 1 GigE

•  For MR-SKEW, IPoIB (32Gbps) gives up to 15% over 10 GigE

BPOE-5 2014

Performance Evaluation with Varying Key/Value Pair Sizes

0 50

100 150 200 250 300 350 400

4 8 16 32

Job

Exe

cutio

n T

ime(

seco

nds)



0

50

100

150

200

250

300

4 8 16 32

Job

Exe

cutio

n T

ime(

seco

nds)



28

•  MR-AVG on Cluster A, on 4 node cluster, 16 maps and 8 reduces •  For both key/value pair sizes, IPoIB (32Gbps) gives up to

•  22-24% improvement over 1 GigE and •  10-13% improvement over 10 GigE

•  Increasing key/value pair sizes lowers job execution time for a given shuffle data size

MR-‐AVG micro-‐benchmark with 100 B key/value size MR-‐AVG micro-‐benchmark with 10 KB key/value size

BPOE-5 2014

Performance Evaluation with Varying Number of Map and Reduce Tasks

•  IPoIB (32Gbps) gives better improvement with increasing concurrency For instance –  up to 32% for IPoIB (32Gbps) for 32 GB shuffle data size –  up to 24% for 10GigE for 32 GB shuffle data size

0

100

200

300

400

500

4 8 16 32

Job

Exec

utio

n Ti

me

(sec

onds

)

Shuffle Data Size (GB)

10GigE-4M-2R10GigE-8M-4R

IPoIB QDR-4M-2RIPoIB QDR-8M-4R

MR-‐AVG with different number of maps and reduces

29

•  MR-AVG on Cluster A, on 4 node cluster, 16 maps and 8 reduces, 1 KB key/value pair size

•  Performance evaluations with –  8 map and 4 reduce tasks (8M-4R) –  4 map and 2 reduce tasks (4M-2R)

•  IPoIB (32Gbps) outperforms 10 GigE by about 13% on an average (over different data sizes)

BPOE-5 2014

Performance Evaluation with Varying Number of Map and Reduce Tasks

•  IPoIB (32Gbps) gives better improvement with increasing concurrency For instance –  up to 32% for IPoIB (32Gbps) for 32 GB shuffle data size –  up to 24% for 10GigE for 32 GB shuffle data size

0

100

200

300

400

500

4 8 16 32

Job

Exec

utio

n Ti

me

(sec

onds

)


10GigE-4M-2R10GigE-8M-4R

IPoIB QDR-4M-2RIPoIB QDR-8M-4R

MR-‐AVG with different number of maps and reduces

30

•  MR-AVG on Cluster A, on 4 node cluster, 16 maps and 8 reduces, 1 KB key/value pair size

•  Performance evaluations with –  8 map and 4 reduce tasks (8M-4R) –  4 map and 2 reduce tasks (4M-2R)

•  IPoIB (32Gbps) outperforms 10 GigE by about 13% on an average (over different data sizes)

BPOE-5 2014

Performance Evaluation with Different Data Types

0 200 400 600 800

1000 1200 1400 1600 1800

4 8 16 32 64

Job

Exec

utio

n Ti

me

(sec

onds

)


1GigE10GigE

IPoIB (32Gbps)

0 200 400 600 800

1000 1200 1400 1600 1800

4 8 16 32 64

Job

Exec

utio

n Ti

me

(sec

onds

)


1GigE10GigE

IPoIB (32Gbps)

31

•  MR-RAND on Cluster A, key/value pair size of 1 KB on 4 node cluster, 16 maps and 8 reducers

•  For both BytesWritable and Text Data Type, IPoIB (32Gbps) shows,

•  significant improvement potential for larger data sizes over 10 GigE on an average

•  similar improvement potential to both types

MR-‐RAND micro-‐benchmark with BytesWritable MR-‐RAND micro-‐benchmark with Text

BPOE-5 2014

Resource Utilization Analysis

0

20

40

60

80

100

0 50 100

CPU

Util

izat

ion(

%)

Sampling Point

1GigE10GigE

IPoIB (32Gbps)

0

200

400

600

800

1000

0 50 100

Netw

ork

Thro

ughp

ut (M

B/s)

Sampling Point

1GigE10GigE

IPoIB (32Gbps)

32

•  MR-AVG on Cluster A, intermediate data size of 16 GB, key/value pair size of 1 KB on 4 node cluster,16 maps and 8 reduces

•  IPoIB (32Gbps) gives best peak bandwidth at 950 MB/s •  CPU trends are similar for different interconnects

CPU uPlizaPon one one slave node Network Throughput on one slave node

BPOE-5 2014

Outline





•  Conclusion &Future Work

33

BPOE-5 2014

Performance of IPoIB Vs. RDMA – MR-AVG

•  Cluster B, key/value pair size of 1 KB, 32 maps and 16 reducers, on 8 and 16 node cluster

•  For MR-AVG, RDMA-based MapReduce over over IPoIB (56Gbps) gives an improvement of

–  Up to 30% for 8 slave nodes (128 cores) –  Up to 28% for 16 slave nodes (256 cores)

0

100

200

300

400

500

600

700

800

8 16 32 64

Job

Exe

cutio

n T

ime(

seco

nds)


IPoIB (56Gbps) RDMA (56Gbps)

0

100

200

300

400

500

600

700

16 32 64 128

Job

Exe

cutio

n T

ime(

seco

nds)



MR-‐AVG with 8 slave nodes on Cluster B MR-‐AVG with 16 slave nodes on Cluster B

34

BPOE-5 2014

Performance of IPoIB Vs. RDMA – MR-RAND

•  For MR-RAND, RDMA-based MapReduce over IPoIB (56Gbps) gives an improvement of –  Up to 28% for 8 slave nodes (128 cores) –  Up to 25% for 16 slave nodes (256 cores)

•  Illustration performance improvement obtained by leveraging RDMA for shuffle and sort phase of MapReduce

0

100

200

300

400

500

600

700

800

8 16 32 64

Job

Exe

cutio

n T

ime(

seco

nds)



0

100

200

300

400

500

600

700

16 32 64 128

Job

Exe

cutio

n T

ime(

seco

nds)



MR-‐RAND with 8 slave nodes on Cluster B MR-‐RAND with 16 slave nodes on Cluster B

35

BPOE-5 2014

Outline






36

BPOE-5 2014

Conclusion and Future Work •  Study of factors that can significantly impact performance

of Hadoop MapReduce (like network protocol etc.)

•  Design, development and implementation of a micro- benchmark suite for stand-alone MapReduce –  Help developers enhance their MapReduce designs –  Support for both Hadoop 1.x and 2.x

•  Performance evaluations with our micro-benchmark suite over different interconnects on modern clusters

•  Future Work –  Enhance micro-benchmark suite to real-world workloads –  Investigate more data types –  Make the micro-benchmarks publicly available through the HiBD

project (OHB) 37

BPOE-5 2014

Thank You! {shankard, luxi, rahmanmd, islamn, panda}

@cse.ohio-‐state.edu

Network-‐Based Compu$ng Laboratory h\p://nowlab.cse.ohio-‐state.edu/

The High-‐Performance Big Data Project h\p://hibd.cse.ohio-‐state.edu/

38

A Micro-Benchmark Suite for Evaluating Hadoop MapReduce …...Hadoop MapReduce 5 Performance of...

Documents

Transcript of A Micro-Benchmark Suite for Evaluating Hadoop MapReduce …...Hadoop MapReduce 5 Performance of...